exploring nonlinearities in financial systemic risk · exploring nonlinearities in financial...

Exploring Nonlinearities in Financial Systemic Risk

Marcin Wolski1,2

1Center for Nonlinear Dynamics in Economics and Finance (CeNDEF),

University of Amsterdam, Valckenierstraat 65-67, 1018 XE Amsterdam, NL2Bielefeld Graduate School of Economics and Management (BiGSEM),

Bielefeld University, Universitatsstraße 25, D-33615 Bielefeld, DE

November 2013

Abstract

We propose a new methodology of assessing the effects of individual institution’s risk

on the others and on the system as a whole. We build upon the Conditional Value-

at-Risk approach, however, we introduce the explicit Granger causal linkages and we

account for possible nonlinearities in the financial time series. Conditional Value-at-Risk-

Nonlinear Granger Causality, or NCoVaR as we call it, has regular asymptotic properties

which makes it particularly appealing for practical applications. We test our approach

empirically and assess the contribution of the euro area financial companies to the overall

systemic risk. We find that only a few financial institutions pose a serious ex ante threat

to the systemic risk, whereas, given that the system is already in trouble, there are more

institutions which hamper its recovery. Moreover, we discover non-negligible nonlinear

structures in the systemic risk profile of the euro zone.

JEL Codes: C1, G01, G20, G32

Keywords: Systemic Risk, Financial Risk, Nonparametric Methods

1 Introduction

The 2007-2009 crisis shed new light on the complexity within the financial sector. The linkages

and risk exposures between various institutions proved to be of great significance in trans-

mitting distress across the whole financial system. Additionally, during systemic events the

The author would like to thank Cees Diks and participants of the seminars at the University of Amsterdamfor their useful comments. Numerical simulations, reported in this paper, were run on the Lisa ComputeCluster. The author acknowledges also the financial support from the European Doctorate in Economics -Erasmus Mundus. The corresponding e-mail addresses is [email protected].

1

malaise spreads across the financial world rapidly through indirect channels, like price effects or

liquidity spirals (Brunnermeier, 2009). In effect, market values of various financial assets tend

to move closer together, drifting away from their fundamentals. In particular, one observes

high regularities in their tail co-movements (Adrian and Brunnermeier, 2011).

Because of its strong adverse effects on the real economy, great attention has been paid

to measuring and monitoring systemic risk, i.e. risk that the entire financial system collapses,

and individual risk exposures. The majority of econometric approaches in these fields focus

on co-risk measures, where the risk of the financial system is assessed in relation to the risk of

individual institutions. The intuition behind these models lies in negative externalities which

one institution imposes on the others and on the system as a whole. As argued by Adrian

and Brunnermeier (2011), these externalities are a consequence of excessive risk taking and

leverage. Given, for instance, that one institution is facing a liquidity shock, it liquidates its

assets at fire-sale prices as given, affecting borrowing constraints of others and actually causing

the fire-sale prices. A wonderful summary of research in this field might be found in Acharya

(2009), Acharya et al. (2010) or Adrian and Brunnermeier (2011).

A commonly used econometric approach, in the growing body of literature on this topic,

is Conditional Value-at-Risk (CoVaR), attributed to Adrian and Brunnermeier (2011). It is

built around the concept of Value-at-Risk (VaR), which determines the maximum loss on

returns within the γ-percentile confidence interval (Kupiec, 2002). CoVaR assesses VaRγ of

one institution conditional on a distress in the other. In particular, given that the former

represents the system, one may associate CoVaR with the systemic risk measure.

A clear shortcoming of such an approach lies in its susceptibility to model misspecifica-

tion. Imagine, that returns come from an unknown probability distribution F , with density

f . Assume now that f is steeper or nonlinear around its VaRγ . Clearly, standard paramet-

ric approaches oversee this irregularity so that even a small variation in VaRγ might affect

co-risk results. In this paper we develop a methodology which corrects for this shortcoming,

contributing to the discussion on nonlinear economic dynamics in systemic risk.

The existence of nonlinearities in the field has been already recognized. Huang et al.

(2010) suggest that a bank’s contribution to the systemic risk is roughly linear in its default

probability and highly nonlinear with respect to institution size and asset correlation. This is

supported by empirical observations of the financial markets in He and Krishnamurthy (2012).

In fact, He and Krishnamurthy (2012) build a theoretical model which matches nonlinear

dynamics across different economic variables, including systemic risk. XiaoHua and Shiying

(2012) investigate the topic from the neural network perspective and design the early warning

mechanism accordingly. This paper aims to propose a formal approach to assess the relevance

of nonlinearities in driving systemic events.

We build our approach around the intuition of CoVaR. In particular, we focus on the

Granger causal effect that a distress in one institution changes to a distress in the other or in

the whole system, where a distress is defined by VaRγ .

2

There are two main novelties in our methodology. The first one is the notion of causality.

The basic CoVaR does not distinguish between direct causal and common factor effects. Adrian

and Brunnermeier (2011) treat it as a virtue rather than a problem, arguing that common factor

effects are of more importance when dealing with systemic risk, which shall be particularly true

for the herding behavior (Brunnermeier et al., 2009). One may, however, want to study the

causal relations explicitly. Imagine for instance a group of the biggest financial institutions.

Since they do not only trade with each other but also serve as clearing houses or liquidity

backstops for smaller parties, they are central to the financial system. Now, imagine that one

of them is in trouble. It affects all the banks which are exposed to its risk, but since it is

relatively large its distress might alone translate into problems in the entire financial system.

The causal kind of reasoning seems therefore particularly appealing for policy makers and

central bankers who in fact might want to focus on preventing this individual causal relation.

Another justification for considering causality in individual and systemic risk lies in its

possible applications to networks and contagion analysis (see for instance Chinazzi and Fagiolo

(2013)). Looking at any pair of institutions, the possible risk effects of one on another do not

have to be bilaterally equal (as they are assumed to be in a non-causal setting). For instance,

a lender has different kind of risk exposure to a creditor than the other way around. Causality

captures that phenomenon explicitly allowing for a more detailed analysis on network spillovers,

cascades and shock propagation.

In our study we employ the causality of Granger type (Granger, 1969), as it is intuitive

and does not bring many model restrictions. It has been also successfully applied as a network

mapping tool in financial analysis (Gao and Ren, 2013).

The second novelty lies in the definition of financial distress. In our study we assume that

an institution is in trouble when it is around its VaRγ . Practically speaking, our definition

captures the majority of events which fall below VaRγ together with some of the events above

it. The reason why we allow for some variation around VaRγ lies in its possible nonlinear

structure, whose role we want to study explicitly. We recognize that our definition might not

capture some of the extreme values from the left tail of the distribution, being potentially

susceptible to black swans (Taleb, 2010). Our analysis shows, however, that the optimal region

around VaRγ is very slowly decreasing with the sample size, somehow hampering the risk of

neglecting the extreme events.

In our analysis we consider two scenarios of potential Granger causality. In the first setting

we investigate the role of individual institutions in blocking the recovery of the system which

is already under distress. In the second scenario we measure the contribution of individual

institutions to the systemic troubles. The second setting is more similar to the standard

understanding of systemic risk (Acharya, 2009) and might be useful in ex ante applications.

The first scenario might be perceived either as a kind of a robustness check or a policy relevant

tool for ex post actions. Indeed, if the system is already in trouble one may want to determine

which of its parts are hampering its recovery. In fact, we could think of these two scenarios from

3

a perspective of a doctor who prescribes precautionary drugs or is trying to heal an already

sick patient.

The paper is organized as follows. In Section 2 we explain the methodology of Conditional

Value-at-Risk-Nonlinear Granger Causality (or NCoVaR for simplicity). We evaluate asymp-

totic properties of the test statistic and we confirm them numerically. In Section 3 we apply

our approach to the euro zone financial sector and evaluate which institutions got the most

significant impact on the systemic risk in years 2000-2012. Section 4 concludes.

2 Methodology of NCoVaR

Let us first bring some intuition behind the Conditional Value-at-Risk and Granger causal-

ity separately and then combine them to build CoVaR-NGraCo (Conditional Value-at-Risk-

Nonlinear Granger Causality) or NCoVaR for simplicity. In the standard setting we consider

two institutions, i and j, whose returns on assets are given by Xi and Xj , respectively. Talking

about systemic risk, we set j to be some aggregate variable so that we investigate the relation-

ship between institution i and the system as a whole. Following the original CoVaR literature,

let us define VaRγ as the left γ-quantile of the returns of a given institution. (In practice γ is

chosen from {0.01, 0.05, 0.1}.) For institution i we have therefore

P (Xi ≤ V aRiγ) = γ, (1)

or equivalently

V aRiγ = inf{xi : FXi(xi) ≥ γ}, (2)

where FXi is the cumulative distribution function of Xi. (For institution j, the notation

is analogous throughout the paper.) The intuition behind CoVaR is to evaluate VaRγ of

institution j conditional on some event associated with institution i. In particular, Adrian and

Brunnermeier (2011) consider two conditioning events, i.e. institution i is at its VaRiγ or at its

median (VaRiγ=0.5 = Mediani). By comparing the difference between the two, it is possible to

estimate the risk contribution of institution i onto j, denoted by ∆CoVaR.

In our study we follow similar reasoning as in Adrian and Brunnermeier (2011), however,

we add a (discrete) time dimension. For any period t, let us define the future returns by GXit ,

and past and/or current returns by FXit . Following Granger (1969), we say that returns of

institution i are Granger causing those of institution j if FXit contains additional information

on GXjt which is not already contained in FXj

t alone. We formulate the definition of conditional

Granger causality analogously, i.e. we say that returns of institution i are Granger causing

those of institution j if, conditional on some past or current events A(FXi

t

)and B

(FXj

t

),

FXit contains additional information on GXj

t which is not already contained in FXjt alone.

Given the intuition behind the CoVaR and conditional Granger causality, we may now turn

4

to NCoVaR. Similarly to ∆CoVaR, we test the difference in Granger causal risk effects from

institution i on j, between two conditioning events, i.e. when institution i is or was in trouble

(or around its VaRiγ) and when it is and was around the median of its returns. An advantage

of allowing institutions to be around (and not exactly at) their VaRγ or median levels is that

we could thereof account for possible nonlinearities in corresponding distributions - something

the original methodology could not capture. In particular, we consider µ-radius ball (µ > 0)

centered at VaRγ or median. (The following reasoning holds for G and F being multivariate,

provided that VaRγ and medians are taken over the marginals.) We allow also for conditioning

on the past and/or current realizations of Xjt . To put it formally, let us formulate the following

definition of NCoVaR.

Definition 1. Given any stationary bivariate process {(Xit , X

jt )}, we say that {Xi

t} is a non-

linear CoVaR Granger cause of {Xjt } if

P(GXj

t

µ≈VaRj

γ |FXit

µ≈VaRi

γ ,B(FXj

t

))6= P

(GXj

t

µ≈VaRj

γ |FXit

µ≈Mediani,B

(FXj

t

)),

where ≈µ represents a µ-radius ball, G denotes future and F denotes past and/or current states

of the corresponding variables and B(.) reflects some event over the argument.

In this study, we consider two possible scenarios. In the first one, we assume that institution

j is already in distress, so that potential Granger causal risk effects from institution i do not

only induce even higher losses on j but also can clog its recovery. The second scenario is more

similar to the traditional risk analysis, where future troubles in institution j come directly from

the past problems of institution j. One may thereof reformulate Def. 1 in form of possible two

scenarios, which we investigate throughout this paper.

Scenario 1. Given any stationary bivariate process {(Xit , X


t} is a non-

linear CoVaR Granger cause of {Xjt } in tail if

P(GXj

t

µ≈VaRj

γ |FXit

µ≈VaRi

γ ,FXjt

µ≈VaRj

γ

)6=

P(GXj

t

µ≈VaRj

γ |FXit

µ≈Mediani,FXj

t

µ≈VaRj

γ

),


of the corresponding variables.

Scenario 2. Given any stationary bivariate process {(Xit , X


t} is a non-

linear CoVaR Granger cause of {Xjt } in median if

P(GXj

t

µ≈VaRj

γ |FXit

µ≈VaRi

γ ,FXjt

µ≈Medianj

)6=

P(GXj

t

µ≈VaRj

γ |FXit

µ≈Mediani,FXj

t

µ≈Medianj

),


5

of the corresponding variables.

In practice it is impossible to condition of the infinite sets of future or past realizations

of variables of interest. Therefore, we reformulate G and F as finite sets of future periods or

lags, respectively. We limit ourselves to the canonical setting where GXjt = Xj

t+1, as it is most

commonly used in practical Granger causality testing, however, our reasoning holds for any

GXjt = Xj

t+k, 1 ≤ k < ∞. Similarly, we replace FXit and FXj

t by Xit,li

= {Xit−li+1, . . . , X

it}

and Xjt,lj

= {Xjt−lj+1, . . . , X

jt }, where li ≥ 1, lj ≥ 1 denote the number of lags of a correspond-

ing variable.

In Granger causality testing, the goal is to find evidence against the null hypothesis of no

causality, which according to Def. 1 is represented by equivalence in conditional probability.

For a strictly stationary process {(Xit , X

jt )}, the null is the statement about the invariant distri-

bution in conditional VaRγ levels of the (li+ lj + 1) -dimensional vector Wt = (Zt, Xit,li, Xj

t,lj),

where we substitute Zt = Xjt+1. (For clarity purposes and to bring forward the fact that we

consider the invariant distribution of Wt, we drop the time index, so that W = (Z,Xi, Xj).) It

is helpful to restate the problem in terms of ratios of joint densities evaluated at given quantiles

as under the null the density of Z evaluated around its VaRγ level and conditional on specific

events in Xi and Xj is equal the same density conditional on the different set of events in Xi

and Xj . Therefore, the joint probability density function, together with its marginals must

satisfy

fZ,Xi,Xj

(zγ , x

iγ , x

j∗

)fXi,Xj

(xiγ , x

j∗

) =fZ,Xi,Xj

(zγ , x

im, x

j∗

)fXi,Xj

(xim, x

j∗

) , (3)

where zγ = VaRZγ , xiγ = VaRi

γ , xim = Mediani and ∗ distinguishes between Scenario 1 and 2

as xjγ = VaRjγ or xjm = Medianj , respectively. Since Eq. (3) holds for any quantile of the

vector (Z,Xi, Xj) in the support of Z,Xi, Xj , Eq. (3) might be equivalently rewritten as

fZ,Xi,Xj

(zγ , x

iγ , x

j∗

)fXi,Xj

(xim, x

j∗

) =fXi,Xj

(xiγ , x

j∗

)fXi,Xj

(xim, x

j∗

) fZ,Xi,Xj

(zγ , x

im, x

j∗

)fXi,Xj

(xim, x

j∗

) . (4)

Natural methodology to assess Eq. (4) comes from the Baeck and Brock (1992) test for condi-

tional independence, or its Granger causal version from Hiemstra and Jones (1994). However,

as showed by Diks and Panchenko (2005) and Diks and Panchenko (2006), these tests can

severely over-reject in Granger causal setting, because its dependence on the conditional vari-

ance. Diks and Panchenko (2006) propose to add a positive weight function g(z, xi, xj) and,

given that the null should hold in the support of the joint densities, it might be equivalently

6

written as

τg ≡

fZ,Xi,Xj

(zγ , x

iγ , x

j∗

)fXi,Xj

(xim, x

j∗

)−

fXi,Xj

(xiγ , x

j∗

)fXi,Xj

(xim, x

j∗

) fZ,Xi,Xj

(zγ , x

im, x

j∗

)fXi,Xj

(xim, x

j∗

) g(z, xi, xj) = 0.

(5)

Diks and Panchenko (2006) discuss several possibilities of choosing g(z, xi, xj). In this study we

focus on g(z, xi, xj) = fXi,Xj

(xim, x

j∗

)2, as the estimator of τg has a corresponding U-statistic

representation, bringing the desired asymptotic normality properties for weakly dependent

data. Substituting into Eq. (5), one finds that

τ = fZ,Xi,Xj

(zγ , x

iγ , x

j∗)fXi,Xj

(xim, x

j∗)− fXi,Xj

(xiγ , x

j∗)fZ,Xi,Xj

(zγ , x

im, x

j∗). (6)

To evaluate the data driven representation of τ , we rely on kernel methods. In particular,

we consider the local density estimator

fW (w) =ε−dW

n

n∑k=1

K

(w − wk

ε

), (7)

where n is the sample size, ε is the bandwidth parameter (similar to µ from the Def. 1), d

reflects the dimensionality of a given vector W and K(.) is a bounded Borel function RdW → Rsatisfying∫

|K(t)|dt <∞,∫K(t)dt = 1 and |tK(t)| → 0 as |t| → ∞. (8)

In practice, K(.) is often chosen to be a probability density function (Wand and Jones, 1995).

In order to guarantee the consistency of the pointwise density estimators, we assume that the

bandwidth parameter ε comes from the sequence εn, which is slowly decreasing with the sample

size, i.e.

εn → 0 and nεn →∞ as n→∞. (9)

Parzen (1962) shows that under conditions (8) and (9) and provided that f is continuous at

w, the estimate of density f at a given point w is consistent.

7

Given a given bandwidth ε, a natural estimator for τ is found to be

Tn(ε) = C

n∑k=1

n∑p=1

[K

((zγ , x

iγ , x

j∗)T − (zk, x

ik, x

jk)T

ε

)K

((xim, x

j∗)T − (xip, x

jp)T

ε

)

− K

((xiγ , x

j∗)T − (xik, x

jk)T

ε

)K

((zγ , x

im, x

j∗)T − (zp, x

ip, x

jp)T

ε

)],

(10)

with

C =ε−dZ−2dXi−2dXj

n2. (11)

The asymptotic distribution of the test statistic might be derived from the behavior of the

properties of the second order U-statistic, as described in Serfling (1980) and van der Vaart

(1998).

Theorem 1. Under conditions (8) and (9), for a given VaRγ levels and given bandwidth

parameter sequence εn, test statistic Tn(εn) satisfies:

√n

(Tn(εn)− τ)

Sn

d−→N (0, 1),

where Sn is the autocorrelation consistent estimator of the asymptotic standard deviation of√n(Tn(εn)− τ).

The proof of Theorem 1 might be found in A. As argued by Diks and Panchenko (2006),

although the test statistic is not positive definite, the one-sided test, i.e. rejecting on larger

values, turns out to yield better performance.

In this study we choose γ to be 0.05 as it is most commonly applied VaR significance

level. We calculate VaRγ from the empirical quantile function (Jones, 1992). We closely follow

the literature on nonparametric Granger causality testing (Hiemstra and Jones, 1994; Diks

and Panchenko, 2006) and take the square kernel function.1 The square kernel form of the

estimator in Eq. (7), can be rewritten as

fSQW (w) =(2ε)−dW

n− 1

n∑k=1

I(||w − wk|| < ε), (12)

where I(||w − wk|| < ε) is the indicator function taking values 1 for any ||w − wk|| < ε and

zero otherwise, and ||.|| is the supremum norm over all the dimensions.

1Up to the scaling factor, the square kernel is parallel to the rectangular (uniform) kernel, with all itsmoments being similar. The asymptotic properties of the test statistic are, however, robust to any kernelspecification, provided that it satisfies conditions (8) and (9).

8

2.1 Optimal bandwidth

Although the asymptotic normality of the test statistic holds for arbitrary decreasing sequence

of bandwidths as long as it satisfies condition (9), it influences the power of the test to a

great extent (Silverman, 1998). Therefore, in order to improve the performance of the test,

we calculate the optimal size of the bandwidth explicitly. Following Wand and Jones (1995)

and Silverman (1998), the optimal bandwidth shall minimize the Mean Squared Error (MSE)

of Tn(εn), which may be decomposed into the sum of variance and squared bias of Tn(εn). In

our inference it is worthwhile to point out that the optimal bandwidth values of Tn(εn) do not

violate the consistency properties of any of the density estimators.

Corollary 1. Under conditions (8) and (9), the MSE optimal sequence of bandwidths of Tn(εn)

guarantee consistency of any of the pointwise density estimators contributing to Tn(εn).

The proof of Corollary 1 is given in B. In fact, the optimum rate of convergence of the

bandwidth of Tn(εn) is slightly faster than these of individual density estimators, but still much

slower than n−1. This is caused by increased variance of a product of two estimators compared

to their individual variances. Therefore, in order to control for this effect, the sequence of

optimal bandwidths of Tn(εn) should decrease at a slightly faster rate as n → ∞, but never

as fast as n−1. In testing for systemic risk this proves to be of large importance as with a

bandwidth parameter decreasing just slightly with the sample size we are still able to capture

the majority of returns which are left to VaRγ .

In evaluating the optimal bandwidth value we rely on the Monte Carlo methods. Correcting

for the weak dependency, we apply the autocorrelation consistent estimator for the variance of

Tn(ε), as proposed in Newey and West (1987). It might be verified that for a given bandwidth

ε, the bias of Tn(ε) might be calculated from the Taylor expansion around any point as

E[Tn(ε)]− τ =1

2κ2ε

2[fZ,Xi,Xj

(zr, x

ir, x

j∗)∇2fXi,Xj

(xis, x

j∗)

+ fXi,Xj

(xis, x

j∗)∇2fZ,Xi,Xj

(zr, x

ir, x

j∗)

− fXi,Xj

(xir, x

j∗)∇2fZ,Xi,Xj

(zr, x

is, x

j∗)

− fZ,Xi,Xj

(zr, x

is, x

j∗)∇2fXi,Xj

(xir, x

j∗)]

+ o(ε2),

(13)

where κ2 is the second moment of the kernel and ∇2fW (w) is the trace of the second derivative

of density evaluated at point w. Up to the error of order o(ε2), Eq. (13) has a plug in estimator,

which might be easily calculated using kernel methods (Wand and Jones, 1995).

To give an example on the size of the optimal bandwidth value, we perform a bootstrap

experiment on the same bivariate process as in the Jeong et al. (2012), i.e.

xit = 1 +1

2xit−1 + r1,t

xjt =1

2xjt−1 + c

(xit−1

)2+ r2,t,

(14)

9

0.4 0.6 0.8 1.0 1.2 1.4

0.00

0000

00.

0000

002

0.00

0000

40.

0000

006

Bandwidth

MS

E

n=100n=200n=500n=1000

(a) Null hypothesis as in Sc. 1.

0.4 0.6 0.8 1.0 1.2 1.40.

0000

000

0.00

0000

20.

0000

004

0.00

0000

6

Bandwidth

MS

E

n=100n=200n=500n=1000

(b) Null hypothesis as in Sc. 2.

Figure 1: MSE of the test statistic for bandwidth values in the range [0.3, 1.5] and for differentsample sizes, aggregated over 1000 simulations.

where r1,t and r2,t independent standard normal variables. The biggest advantage of process in

Eq. (14) is its tuning parameter on Granger causality, c. Clearly, if c = 0 the model corresponds

to the hypothetical scenario of no Granger causality from the first lags of Xit on Xj

t . The larger

parameter c becomes, the stronger Granger causal effect, which we may control for explicitly.

We perform 1000 simulations of normalized data of process from Eq. (14) for different

sample sizes and evaluate MSE of the test statistic for different bandwidth values from the

range [0.3, 1.5]. For practical reasons, the interval between the bins is taken as 0.01. We take

lags of order 1 for both variables, as suggested by the underlying lag structure in Eq. (14).

The results for two scenarios of Granger causality are presented in Fig. 1 and the optimal

bandwidths are reported in Table 1.

It is straightforward to notice the differences of the MSE curves between two settings.

Firstly, for the same sample size and ε, Scenario 2 demonstrates larger MSE than in Scenario

1. Secondly, in Scenario 1 the MSE curve becomes flatter, whereas in Scenario 2 the visible

U-shape is preserved as the sample size increases. These, in fact, are direct consequences of

the curvature of the true distribution around particular quantiles. Scenario 1 is driven by the

tail dependence, where the curvature shall be relatively flatter. On the contrary, Scenario 2

represents the relation between the tail and the median, where the distribution shall be more

bell-shaped or simply steeper. This, in fact, shows up in the steepness and in the relative size

of the MSE curve. As expected, the minimum of the MSE curves is decreasing with the sample

size in both scenarios (see Table 1).

Reported optimal bandwidth values represent the radius around the V aRγ which is being

10

Table 1: Optimal bandwidth values for test statistic evaluated for the process from Eq. (14) fordifferent sample sizes and for two scenarios. The values represent means over 1000 simulations.

n = 100 n = 200 n = 500 n = 1000ε∗ (Sc.1) 0.74 0.66 0.6 0.52ε∗ (Sc.2) 0.68 0.64 0.48 0.44

considered in the NCoVaR. One may readily observe that Scenario 1 has slightly larger band-

widths than Scenario 2. We may view this as a result of scarcity of data in tails compared

with that around the median. Extracting information from tails requires, on average, slightly

larger windows in comparison to the region near the median (Caers and Maes, 1998).

Bootstrapping optimal bandwidths is a powerful technique which might be applied to any

data set without assuming an underlying process structure. We recognize, however, that it

might take a lot of computational time. For very large samples we suggest taking bins of 0.02

or 0.05 in order to make it computationally less demanding. Our simulations confirm that the

power of the test is preserved in the range [ε∗ − 0.05, ε∗ + 0.05].

2.2 Numerical performance

We perform two experiments to evaluate the practical side of the test. In both we rely on

Monte Carlo methods on the example of the process in Eq. (14).2 In the first one, we assess

the distribution of the test statistic under the null, evaluated for different sample sizes for

500 runs. In the second experiment, we estimate the power of the test. Given that the null

hypothesis is violated (c > 0), we estimate rejection rates for different nominal significance

levels. We summarize the results from both experiments in the size-size plots and size-adjusted

power diagrams. The former plots the actual against nominal cumulative rejection rates under

the null, showing the size distortions. The later shows the power of the test corrected for

the possible size bias, plotting the observed cumulative rejection rates under the alternative

(actual power) against observed rejection rates under the null (actual size). Ideally, the power

function shall be 1 for any significance level larger than 0, however, in practice we would like

to observe an increase in the slope at the origin as the sample size grows. Fig. 2 shows the

size-size diagrams whereas the size-adjusted power plots are presented in Figs 3-5.

Fig. 2 suggests that the nominal size distortions are larger in Scenario 2 than in Scenario

1. Additionally, the size-size curves are flatter in Scenario 1 whereas they are more wavy in

Scenario 2. In fact, this is similar to the pattern observed in the MSE (see Fig. 1) and might

be largely attributed to the curvature of the true distribution around particular quantiles.

One may readily observe from Figs 3-5 that the size-adjusted power of the test increases

2One may expect that the numerical size distortions and power of the NCoVaR test could depend on theexact process specification. Eq. (14) offers a simple testing environment, which has been already applied in thequantile testing literature (Jeong et al., 2012). We leave, therefore, the assessment of the NCoVaR numericalperformance on other processes for future investigation.

11

0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Nominal size

Act

ual s

ize

45 degreen=100n=200n=500n=1000


0.2 0.4 0.6 0.8 1.00.

00.

20.

40.

60.

81.

0

Nominal size

Act

ual s

ize

45 degreen=100n=200n=500n=1000


Figure 2: Size-size diagram of the NCoVaR test for the process from Eq. (14) for differentsample sizes over 500 simulations.

with the sample size and with the strength of Granger causality. Nevertheless, there are two

main patterns emerging from the numerical analysis which deserve pointing out.

Firstly, for relatively smaller size the power of the test is higher for Scenario 1 than for

Scenario 2. This is again the result of model dynamics, where the underlying relation on

variable j, i.e. P (Xjt+1 ≈ VaRj

γ |Xjt ≈ Medianj) is more rare to observe on the process from

Eq. (14). Practically speaking, as the sample size gets larger this effect is hampered.

Secondly, the size-adjusted test power is almost negligible for very small Granger causality

and short time series. Clearly, one shall blame the relative scarcity of observations around

quantiles for this discomfort. In order to apply the test to shorter data sets, we propose two

solutions to overcome this issue. The first one comprises different kernel specifications. The

square kernel takes into account only observations which are ε-close to the quantile, leaving

out many possibly informative data points. Replacing kernel by a smoother one, like Gaussian

or logistic, should thereof correct for this effect. The second possible solution lies in improving

the precision of the density estimators. In the standard kernel estimators (like square kernel

estimators applied here) the bias is of order ε2 (Wand and Jones, 1995). Making the bias

smaller should decrease the disinformative effect of the observations around a given quantile so

that keeping the sample size fixed we get relatively better representation of the true Granger

causal relation, which translates into improved test performance.

One may consider Data Sharpening (DS) as being potentially attractive bias reduction

method in our setting. Following Hall and Minnotte (2002), the idea behind DS is to slightly

perturb the original data set in order to obtain desirable estimator properties (here it is the

12

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Actual size

Act

ual p

ower

45 degreen=100n=200n=500n=1000


0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Actual size

Act

ual p

ower

45 degreen=100n=200n=500n=1000


Figure 3: Size-adjusted power for the NCoVaR test for the process from Eq. (14) for c = 0.05for different sample sizes over 500 simulations.

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Actual size

Act

ual p

ower

45 degreen=100n=200n=500n=1000


0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Actual size

Act

ual p

ower

45 degreen=100n=200n=500n=1000



13

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Actual size

Act

ual p

ower

45 degreen=100n=200n=500n=1000


0.0 0.2 0.4 0.6 0.8 1.00.

00.

20.

40.

60.

81.

0

Actual size

Act

ual p

ower

45 degreen=100n=200n=500n=1000



reduced bias). Diks and Wolski (2013) show that, besides reducing the estimator bias, DS

does not affect other asymptotic properties of the test statistic in a similar Granger causality

setting. Therefore, it seems to be a straightforward extension to NCoVaR for shorter samples.

3 Assessing financial systemic risk

In our analysis we focus on the NCoVaR of individual institutions on the overall systemic

risk. Thereof, we set that j represents the system variable and i is the individual financial

institution.

We approximate the returns on assets by equity returns and take into account financial in-

stitutions publicly traded within the euro zone. In order to make the analysis more transparent

we focus on the companies which build the Euro STOXX Financial Index in years 2000-2012.

Our sample covers thus the Great Recession in Europe (2008-12), the financial crisis (2007-

2009) and the sovereign debt crisis (2010-2012). In total we collect daily equity returns for

48 companies (3 financial, 13 insurance, 23 banks and 9 real estate) and one aggregate index.

For each variable we have 3390 observations. The list of companies, together with the coun-

try of origin and their sector might be found in D. The data have been obtained from the

DataStream.

All time series are stationary at the 1% significance level, according to both Phillips-Perron

and Augmented Dickey-Fuller specifications (Phillips and Perron, 1988; Fuller, 1995). We run

the pairwise tests against the null of no NCoVaR between each company and system variable.

14

SOC

STOXX

ALL

ASS

AXA

PAS

ROM

ESP

SAN

COM

DEU

ERS

AEG

ING

KBC

KLE

MED

NAT

POH

UNI

VIE

WEN

WER

(a) Euro area NCoVaR in Sc. 1.

ERS

STOXX

AEG

AGE

BIR

SON

POPACK

ICA

IMM

UNI

WER

INT

(b) Euro area NCoVaR in Sc. 2.

Figure 6: NCoVaR between euro area individual financial companies and system variable forraw data.

In order to make sure that all the Granger causal relations are nonlinear, we run the same

test specification on VAR-filtered residuals also. In each run the number of lags is taken in

accordance with the Schwarz-Bayes Information Criterion from the VAR specification and the

optimal bandwidth value is approximated by bootstrap. As a robustness check, we also correct

for possible causality in second moments, as suggested in Francis et al. (2010), by running

NCoVaR test on residuals from Dynamics Conditional Correlation GARCH model (Engle,

2002).

The detailed results can be found in D (Tables D.2, D.3 and D.4), however, for presenta-

tional clarity we refer to the star-graphs, which show the NCoVaR between each company and

system as a whole. Width of the arrows represents the inverse of the statistical significance

level of NCoVaR (the stronger the NCoVaR effect, the wider (and darker) the arrow). Fig. 6

shows the results for the raw time series, Fig. 7 depicts the VAR-filtered returns and Fig. 8

refers to the GARCH residuals.

Considering that at least one NCoVaR relation denotes a systemically important institu-

tion, our analysis suggests that out of 48 companies 33 of them might be so described. The

group consists of 3 financial services companies, 6 insurance firms, 19 banks and 5 real estate

companies. In fact, all of the financial services companies in our sample prove to be systemically

important.

There are two main patterns emerging from our analysis. Firstly, there are fewer systemi-

cally risky institutions in Scenario 2. Secondly, NCoVaR in Scenario 1 is on average stronger

than in Scenario 2. These findings hold for the original and VAR- and GARCH-filtered data.

15

ACK

STOXX

AEG AXA

PAS

SON

ROM

BBV

ESS

ESP

SAN

BNP

WEN

DEU

GEC

SOC ING

KBC

MED

POH

UNI

KLE

MAP

COM


GBL

STOXXAGE

BIR

PAS

IMM

POP

ERS

ACK

ICA

VIE

INT

ESS

UNI

KLE


Figure 7: NCoVaR between euro area individual financial companies and system variable forVAR-filtered data.

DEU

STOXX

AGE

ALL

ASS

AXAPAS

ROM

BBV

ESP

SAN

COM

MAP

AEG

ERS

SOC

ING

INT

KBC

UNI

MED

POH

WER

VIE

COR

WEN


ERS

STOXX

ICA

BIR

SON

POP

ACK

UNI

IMM

WER

AEG

MAP


Figure 8: NCoVaR between euro area individual financial companies and system variable forGARCH-filtered data.

16

Interestingly, our study suggests that only a few financial institutions pose a serious ex ante

threat to the systemic risk in the euro area, whereas, given that the system is already in trouble,

there are more institutions which hamper its recovery. This result confirms a common view

in the literature on macro-prudential supervision (Acharya, 2009) that the relative preventive

costs are smaller than those after the crisis has already erupted.

The analysis confirms the nonlinear structure of the institutional contribution to the sys-

temic risk. Filtering out the linear relations and second moment influence does not remove the

co-risk relations among individual companies and system as a whole. Interestingly, after filter-

ing we observe some new co-risk relations emerging. To illustrate this better let us consider

ACK (Ackermans & Van Haaren). The raw data do not show any NCoVaR, however, after

linear filtering it poses a very strong threat to the system’s recovery (see Table D.3 in D; test

statistic of order 6.351 in Scenario 1) and after GARCH filtering it has a weak ex ante effect

on the system’s risk (test statistic of order 1.329 in Scenario 2). One may speculate that there

are some strong purely nonlinear and second moment co-risk effects from ACK on the system

variable, which are being partly offset by their linear equivalents. In other words, under normal

circumstances ACK does not seem to be an important systemic risk contributor. However, in

abnormal times, like a crisis, it reveals its systemic importance.

There is one more finding which we believe is worth pointing out. We confront our re-

sults with the official list of Global Systemically Important Banks (G-SIBs), published by the

Financial Stability Board (FSB) in 2011.3 The FSB recognizes 11 G-SIBs in the euro area.

Our sample covers 8 of them, i.e. Banco Bilbao Vizcaya Argentaria (BBV), Banco Santander

(SAN), BNP Paribas (BNP), Commerzbank (COM), Deutsche Bank (DEU), Societe Generale

(SOC) , UniCredit (Uni) and ING Bank (ING), as a part of the ING Groep. For all of them

we confirm their G-SIB status in at least one NCoVaR setting.

4 Conclusions and discussion

Conditional Value-at-Risk-Nonlinear Granger Causality, or NCoVaR, is a new methodology of

assessing co-risk relations, designed to capture their possible nonlinear Granger causal effects.

Our approach distinguishes between two possible scenarios. In the first one, we test what is the

role of individual institutions in hampering the recovery of others, given that they are already

in distress. In the second scenario, we assess the contribution of individuals to the others’

troubles. We derive the regular asymptotic properties of the NCoVaR test for both scenarios

and we confirm them numerically.

We apply our methodology to assess the systemic importance of financial institutions in

the euro area. Our findings suggest that (i) only a few financial institutions pose a serious ex

ante threat to the systemic risk, whereas, given that the system is already in trouble, there are

3The G-SIBs list is being often updated. In our comparison we consider the most recent version of the list,published on November 11th, 2013.

17

more institutions which hamper its recovery and (ii) there are intriguing nonlinear structures

in its systemic risk profile.

Our study suggests that the most systemically risky institution in our sample is UNI (Uni-

Credit), an Italian bank. In all settings it demonstrates very strong NCoVaR relation to the

system. In 2011 it was recognized by the FSB as G-SIB. This analysis confirms its systemical

importance, revealing also its nonlinear nature. Interestingly, there are two more companies

which demonstrate very strong NCoVaR in 5 out of 6 settings, i.e. ERS (Erste Group Bank),

an Austrian bank, and AEG (Aegon), a Dutch insurer. Only the latter was recognized by the

FSB to be potentially systemically important, with no official view on the former. However,

the former was recognized as systemically important bank for the Austrian financial sector (von

Kruechten et al., 2009). Our results point to potential systemic importance of Erste Group

Bank in the entire euro area.

NCoVaR might be of great use for macro-prudential policy, however, it has to be tested

on other samples and in other periods. It reveals some intriguing phenomena in the co-risk

relations. In order to understand them better, a tempting idea is to investigate the underlying

nonlinear structures analytically in models of aggregate economy. One may also apply NCoVaR

as a mapping tool and bring the risk analysis to the network level.

18

References

V. Acharya. A theory of systemic risk and design of prudential bank regulation. Journal of

Financial Stability, 5(3):224–255, 2009.

V. Acharya, L. Pedersen, T. Philippon, and M. Richardson. Measuring systemic risk. Working

Paper 1002, New York University, 2010.

T. Adrian and M. K. Brunnermeier. CoVaR. Working Paper 17454, FED New York, 2011.

E. Baeck and W. Brock. A nonparametric test for independence of a multivariate time series.

Working Paper 9204, University of Wisconsin – Madison, 1992.

M. Brunnermeier. Deciphering the liquidity and credit crunch 2007-2008. Journal of Economic

Perspectives, 23(1):77–100, 2009.

M. Brunnermeier, A. Crocket, C. Goodhart, A. Perssaud, and H. Shin. The Fundamental

Principles of Financial Regulation: Geneva Reports on the World Economy. Centre for

Economic Policy Research, London, 2009.

J. Caers and M. A. Maes. Identifying tails, bounds and end-points of random variables.

Structural Safety, 20(1):1–23, 1998.

M. Chinazzi and G. Fagiolo. Systemic risk, contagion, and financial networks: A survey. LEM

Working Paper 2013/08, Sant’Anna School of Advanced Studies, 2013.

M. Denker and G. Keller. On U -statistics and v. Mises’ statistics for weakly dependent pro-

cesses. Zeitschrift fur Wahrscheinlichkeitstheorie und verwandte Gebiete, 64:505–522, 1983.

C. Diks and V. Panchenko. A note on the Hiemstra-Jones test for Granger non-causality.

Studies in Nonlinear Dynamics and Econometrics, 9:1–7, 2005. art. 4.

C. Diks and V. Panchenko. A new statistic and practical guidelines for nonparametric granger

causality testing. Journal of Economic Dynamics and Control, 30:1647–1669, 2006.

C. Diks and M. Wolski. Nonlinear granger causality: Guidelines for multivariate analysis.

CeNDEF Working Paper 13-15, University of Amsterdam, 2013.

R. Engle. Dynamic conditional correlation: A simple class of multivariate generalized autore-

gressive conditional heteroskedasticity models. Journal of Business & Economic Statistics,

20(3):339–350, 2002.

B. B. Francis, M. Mougoua, and V. Panchenko. Is there a symmetric nonlinear causal rela-

tionship between large and small firms? Journal of Empirical Finance, 17(1):23–38, 2010.

W. A. Fuller. Introduction to Statistical Time Series (Wiley Series in Probability and Mathe-

matical Statistics). Wiley-Interscience, New York, 1995.

19

B. Gao and R.-e. Ren. The topology of a causal network for the Chinese financial system.

Physica A: Statistical Mechanics and its Applications, 392(13):2965–2976, 2013.

C. W. J. Granger. Investigating causal relations by econometric models and cross-spectral

methods. Econometrica, 37:424–438, 1969.

P. Hall and M. C. Minnotte. High order data sharpening for density estimation. Journal of

the Royal Statistical Society. Series B (Statistical Methodology), 64(1):141–157, 2002.

Z. He and A. Krishnamurthy. A macroeconomic framework for quantifying systemic risk.

Working Paper 233, National Bank of Belgium, 2012.

C. Hiemstra and J. D. Jones. Testing for linear and nonlinear Granger causality in the stock

price – volume relation. Journal of Finance, 49(5):1639–1664, 1994.

X. Huang, H. Zhou, and H. Zhu. Systemic risk contributions. Working Paper 60, Bank for

International Settlements, 2010.

K. Jeong, W. K. Hardle, and S. Song. Nonparametric test for causality in quantile. Econometric

Theory, 28:861–887, 2012.

M. C. Jones. Estimating densities, quantiles, quantile densities and density quantiles. Annals

of the Institute of Statistical Mathematics, 44(4):721–727, 1992.

P. H. Kupiec. Stress testing in a value at risk framework. In M. A. H. Dempster, editor,

Risk Management: Value at Risk and Beyond, pages 76–99. Cambridge University Press,

Cambridge, 2002.

W. Newey and K. D. West. A simple, positive semi-definite, heteroskedasticity and autocorre-

lation consistent covariance matrix. Econometrica, 55(3):703–08, 1987.

E. Parzen. On estimation of a probability density function and mode. The Annals of Mathe-

matical Statistics, 33(3):1065–1076, 1962.

P. C. B. Phillips and P. Perron. Testing for a unit root in time series regression. Biometrika,

75(2):335–346, 1988.

R. J. Serfling. Approximation Theorems of Mathematical Statistics. Wiley, New York, 1980.

B. W. Silverman. Density Estimation for Statistics and Data Analysis (Chapman & Hall/CRC

Monographs on Statistics & Applied Probability). Chapman and Hall/CRC, London, 1998.

N. N. Taleb. The Black Swan: The Impact of the Highly Improbable. Random House Trade

Paperbacks, New York, 2010.

A. van der Vaart. Asymptotic Statistics. Cambridge University Press, Cambridge, 1998.

20

V. von Kruechten, E. Romanova, and M. Schmaus. Austria-based erste group bank ag ‘a’

rating affirmed on high systemic importance. Discussion Paper 2013/08, Standard & Poor’s

Rating Direct, 2009.

M. P. Wand and M. C. Jones. Kernel Smoothing (Chapman & Hall/CRC Monographs on

Statistics & Applied Probability). Chapman and Hall/CRC, London, 1995.

D. XiaoHua and K. Shiying. International early warning nonlinear model for securities systemic

risk based on RBFNN using UDM and BVM. Procedia Engineering, 29(0):1378–1386, 2012.

21

A Asymptotic properties of test statistic (Theorem 1)

We first deal with the properties for the independent sample and consider the dependency

later. By symmetrization with respect to two indices, the test statistic in Eq. (10) has a

corresponding U-statistic representation of the form

Tn(εn) ≡ Tn(ε) =1(n2

) n∑k=1

∑p≤k

K(Wk,Wm), (15)

with Wk = (Zk, Xik,li, Xj

k,lj), k = 1, . . . , n and kernel given by

K(Wk,Wp) =ε−dZ−2dXi−2dXj (n− 1)

2n

[Kk(zγ , x

iγ , x

j∗)Kp(x

im, x

j∗)

− Kk(xiγ , xj∗)Kp(zγ , x

im, x

j∗) +Kp(zγ , x

iγ , x

j∗)Kk(xim, x

j∗)

− Kp(xiγ , x

j∗)Kk(zγ , x

im, x

j∗)],

where for clarity we denote Kk(w) = K((w − wk)/ε) and dZ , dXi and dXj are general repre-

sentations of the dimensionality of G and F operators for particular variables. It is worth to

remind here that subscript n in the test statistic refers to its sequence.

The asymptotic properties of the sequence of test statistic can be derived by the projection

method (van der Vaart, 1998). From the Hayek’s projection lemma we know that the projection

of Tn(ε)− τ on the set of all function of the form∑nk=1 κk(Wk) is given by

Tn(ε) =

n∑k=1

E[(Tn(ε)− τ)|Wk] =2

n

n∑k=1

K1(wk), (16)

where

K1(wk) = EWp

[K(wk,Wp))

]− τ. (17)

Projection Tn(ε) is mean zero sequence with variance 4/nVar(K1(W1)). By the Central

Limit Theorem, one may verify that√nTn(ε) converges in distribution to the normal law with

mean 0 and variance given by 4Var(K1(W1)).

Provided that Var(Tn(ε)) → Var(Tn(ε)) as n → ∞, by Slutsky’s lemma, we now observe

that for a given ε and given quantiles of any independent finite-variance process (Zt, Xit,li, Xj

t,lj),

the sequence√n(Tn(ε)− τ − Tn(ε)

)converges in probability to zero as n→∞. What follows,

the sequence√n (Tn(ε)− τ) converges in distribution to N

(0, σ2

), where

σ2 = 4ζ1, (18)

with ζ1 = Cov(K(W1,W2), K(W1,W

′2))

= Var(K1(W1)).

22

A.1 Dependence

Following the reasoning from Denker and Keller (1983), the above asymptotic normality prop-

erties of the test statistic, Tn(ε), hold for a weakly dependent process if we take into account

the covariance between estimators of particular vectors in the asymptotic variance σ2,

σ2 = 4

[ζ1 + 2

n∑t=2

Cov(K1(W1), K1(Wt)

)]. (19)

According to the kernel specification, the estimator for K1(Wk) is given by

K1(Wk) =(2ε)−dZ−2dXi−2dXj

n

n∑p=1

K(Wk,Wp).

The Newey and West (1987) autocorrelation consistent estimator of σ2 is then

S2n =

B∑b=1

Rbωb, (20)

where B = bn1/4c, Rb is the sample covariance function of K1(Wb) given by

Rb =1

n− b

n−b∑a=1

(K1(Wa)− Tn(ε))(K1(Wa+b)− Tn(ε)), (21)

and ωb is the weight function of the form

ωb =

{1, if b = 1

2− 2(b−1)τ , if b > 1.

(22)

For any finite-variance process (Zt, Xit,li, Xj

t,lj), it follows from Denker and Keller (1983) that

√n

(Tn(ε)− τ)

Sn

d−→N (0, 1), (23)

which completes the proof of Theorem 1.

B Optimal bandwidth sequence (Corollary 1)

For a given bandwidth ε, the MSE of the test statistic might be rewritten as as sum of variance

and squared bias (Wand and Jones, 1995), i.e.

MSE[Tn(ε)] = Var(Tn(ε)) + Bias(Tn(ε))2, (24)

23

where Bias(Tn(ε)) might be calculated explicitly from the Taylor expansion as in Eq. (13) and

variance of the test statistic might be represented as 4S2n/n from A.1. Asymptotic covariance

terms tend to zero as n→∞ so that under the null one might find that the asymptotic variance

of Tn(ε) might be decomposed into

Var(Tn(ε)) = Var(fZ,Xi,Xj

(zγ , x

iγ , x

j∗))Var(fXi,Xj

(xim, x

j∗))

+ Var(fXi,Xj

(xiγ , x

j∗))Var(fZ,Xi,Xj

(zγ , x

im, x

j∗))

+ Var(fXi,Xj

(xim, x

j∗))E[fZ,Xi,Xj

(zγ , x

iγ , x

j∗)]2

+ Var(fZ,Xi,Xj

(zγ , x

iγ , x

j∗))E[fXi,Xj

(xim, x

j∗)]2

+ Var(fXi,Xj

(xiγ , x

j∗))E[fZ,Xi,Xj

(zγ , x

im, x

j∗)]2

+ Var(fZ,Xi,Xj

(zγ , x

im, x

j∗))E[fXi,Xj

(xiγ , x

j∗)]2 + o(1).

One may find that variance and bias of individual density estimators are o(n−1ε−dW ) and

o(ε−2), respectively (Silverman, 1998). Therefore, the dominant terms in the asymptotic vari-

ance are of order o(n−1ε−dZ−dXi−dXj−4).

Taking the first order conditions of the MSE of of individual density estimators, one finds

that the optimum rate of convergence of bandwidth parameter is n−1/(dW+4). Doing the

same for our test statistic, we find that this rate is n−1/(dZ+dXi+dXj ). Therefore, for any finite

dimension, the optimal rate of convergence of the Tn(εn) bandwidth is slightly faster than those

of individual density estimators but never as fast as n−1 which would violate condition (9).

Provided that the optimum rate of convergence of the individual estimators is sufficient for the

consistency (Silverman, 1998), the optimum rate of Tn(εn) guarantees consistency as well.

C Data description and results

D Data description and results

The Euro STOXX Financials Index consists originally of 61 entities. However, only 48 of

them cover years 2000-2012 (see Table D.1). For all of them we collect daily equity prices and

calculate their log returns accordingly. Data comes from the DataStream and covers period

01/01/2000 till 12/31/2012. All time series are stationary at 1% significance level, according to

both Phillips-Perron and Augmented Dickey-Fuller specifications (Phillips and Perron, 1988;

Fuller, 1995).

Table D.1: List of all entities used in the empirical analysis.

Company name/Index Symbol Sector Country

1 Euro STOXX Financials STOXX Aggregate Aggregate2 Ackermans & Van Haaren ACK Financial Services BE3 Aegon AEG Insurance NL4 Ageas AGE Insurance NL5 Allianz ALL Insurance DE6 Assicurazioni Generali ASS Insurance IT7 AXA AXA Insurance FR8 Bank Of Ireland BIR Banks IR9 Bankinter BAN Banks ES

10 Banca Monte Dei Paschi PAS Banks IT11 Banca Popolare Di Milano MIL Banks IT12 Banca Popolare Di Sondrio SON Banks IT13 Banca Popolare Emilia Romagna ROM Banks IT14 BBV Argentaria BBV Banks ES15 Banco Comr. Portugues POR Banks PT16 Banco Espirito Santo ESS Banks PT17 Banco Popolare POP Banks IT18 Banco Popular Espanol ESP Banks ES19 Banco Santander SAN Banks ES20 BNP Paribas BNP Banks FR21 CNP Assurances CNP Insurance FR22 Cofinimmo COF Real Estate BE23 Commerzbank COM Banks DE24 Corio COR Real Estate NL25 Deutsche Bank DEU Banks DE26 Erste Group Bank ERS Banks AT27 Fonciere Des Regions FON Real Estate FR28 Gecina GEC Real Estate FR29 GBL New GBL Financial Services BE30 Societe Generale SOC Banks FR31 Hannover Ruck. HAN Insurance DE32 ICADE ICA Real Estate FR33 Immofinanz IMM Real Estate AT34 ING Groep ING Insurance NL35 Intesa Sanpaolo INT Banks IT36 KBC Group KBC Banks BE37 Klepierre KLE Real Estate FR38 Mapfre MAP Insurance ES39 Mediobanca MED Banks IT40 Muenchener Ruck. MUE Insurance DE41 Natixis NAT Banks FR42 Pohjola Pankki POH Banks FI43 Sampo SAM Insurance FI44 SCOR SCO Insurance FR45 Unibail-Rodamco ROD Real Estate FR46 UniCredit UNI Banks IT47 Vienna Insurance Group VIE Insurance AT48 Wendel WEN Financial Services FR49 Wereldhave WER Real Estate NL

Table D.2: NCoVaR from institution i on the system risk in two scenarios in period 01/01/2000till 12/31/2012 for raw returns. Lags determines the optimal number of lags from the VARspecification using the Schwarz-Bayes Information Criterion. Optimal epsilon values calculatedfrom bootstrap. T-val represents the test statistic of NCoVaR from Eq.(10). (*),(**), (***)denotes one-sided p-value statistical significance at 10%, 5% and 1%, respectively.

Scenario 1 Scenario 2Institution i System variable Lags Opt. ε T-val Lags Opt. ε T-val

1 ACK STOXX 1 0.54 0.913 1 0.44 1.329 *2 AEG STOXX 1 0.28 6.359*** 1 0.4 2.855***3 AGE STOXX 4 0.6 0.213 4 0.6 3.516***4 ALL STOXX 1 0.38 1.45 1 0.2 -6.3655 ASS STOXX 1 0.24 6.351*** 1 0.22 -6.3596 AXA STOXX 1 0.26 6.351*** 1 0.2 -3.0357 BIR STOXX 2 0.6 0.585 2 0.6 2.611***8 BAN STOXX 1 0.38 -1.262 1 0.22 -3.4469 PAS STOXX 1 0.32 6.351*** 1 0.2 -4.38210 MIL STOXX 1 0.24 -6.369 1 0.2 -4.47711 SON STOXX 1 0.3 -3.854 1 0.3 4.448***12 ROM STOXX 1 0.48 1.571 * 1 0.26 -5.83313 BBV STOXX 1 0.38 -1.081 1 0.2 -6.36114 POR STOXX 1 0.26 -6.351 1 0.3 -5.37815 ESS STOXX 1 0.26 -6.357 1 0.2 -5.4416 POP STOXX 1 0.32 -1.551 1 0.26 5.359***17 ESP STOXX 1 0.26 6.351*** 1 0.2 -6.36418 SAN STOXX 1 0.2 6.351*** 1 0.42 0.67319 BNP STOXX 1 0.34 -1.947 1 0.2 -3.12120 CNP STOXX 1 0.3 0.78 1 0.2 -6.36721 COF STOXX 1 0.2 -6.351 1 0.2 -6.36522 COM STOXX 1 0.26 6.351*** 1 0.2 -6.3623 COR STOXX 1 0.58 0.411 1 0.22 -6.37524 DEU STOXX 1 0.24 6.352*** 1 0.24 -6.37125 ERS STOXX 1 0.38 1.576 * 1 0.26 3.295***26 FON STOXX 1 0.26 -6.357 1 0.32 0.05227 GEC STOXX 1 0.24 -6.362 1 0.32 -0.38928 GBL STOXX 1 0.3 -6.359 1 0.42 0.08829 SOC STOXX 1 0.24 6.351*** 1 0.22 -6.36430 HAN STOXX 1 0.32 -0.645 1 0.2 -3.91131 ICA STOXX 1 0.24 -1.474 1 0.34 3.846***32 IMM STOXX 2 0.6 -0.316 2 0.6 2.684***33 ING STOXX 1 0.26 6.351*** 1 0.22 -3.73434 INT STOXX 1 0.28 -2.919 1 0.28 4.994***35 KBC STOXX 1 0.26 6.352*** 1 0.2 -6.35936 KLE STOXX 1 0.24 6.351*** 1 0.52 -0.31537 MAP STOXX 1 0.26 -0.444 1 0.22 -6.37138 MED STOXX 1 0.26 6.351*** 1 0.2 -6.36139 MUE STOXX 1 0.28 -6.351 1 0.2 -6.35740 NAT STOXX 1 0.26 1.417 * 1 0.28 -4.93741 POH STOXX 1 0.28 6.352*** 1 0.24 -3.92542 SAM STOXX 1 0.6 -0.026 1 0.28 -4.66643 SCO STOXX 1 0.32 -2.487 1 0.2 -6.36544 ROD STOXX 1 0.52 0.239 1 0.24 -6.37745 UNI STOXX 1 0.28 6.352*** 1 0.3 5.886***46 VIE STOXX 1 0.32 3.747*** 1 0.2 -6.36247 WEN STOXX 1 0.24 4.733*** 1 0.22 -6.36448 WER STOXX 1 0.3 3.34*** 1 0.44 1.572 *

Table D.3: NCoVaR from institution i on the system risk in two scenarios in period 01/01/2000till 12/31/2012 for VAR-filtered returns. Lags determines the optimal number of lags from theVAR specification using the Schwarz-Bayes Information Criterion. Optimal epsilon values cal-culated from bootstrap. T-val represents the test statistic of NCoVaR from Eq.(10). (*),(**),(***) denotes one-sided p-value statistical significance at 10%, 5% and 1%, respectively.


1 ACK STOXX 1 0.26 6.351*** 1 0.44 1.417 *2 AEG STOXX 1 0.24 6.355*** 1 0.2 -6.363 AGE STOXX 1 0.38 0.983 1 0.24 5.506***4 ALL STOXX 1 0.36 0.247 1 0.2 -6.3635 ASS STOXX 1 0.3 -2.589 1 0.22 -6.3616 AXA STOXX 1 0.32 6.353*** 1 0.2 -6.3587 BIR STOXX 1 0.22 -6.362 1 0.32 4.329***8 BAN STOXX 1 0.28 -2.625 1 0.2 -6.369 PAS STOXX 1 0.34 6.35*** 1 0.3 5.469***10 MIL STOXX 1 0.34 -5.756 1 0.2 -3.40711 SON STOXX 1 0.36 4.458*** 1 0.2 -5.4412 ROM STOXX 1 0.34 2.491*** 1 0.22 -5.62713 BBV STOXX 1 0.2 6.35*** 1 0.2 -6.3614 POR STOXX 1 0.32 -6.36 1 0.24 -6.33615 ESS STOXX 1 0.36 1.998*** 1 0.24 5.202***16 POP STOXX 1 0.34 -3.085 1 0.3 4.667***17 ESP STOXX 1 0.36 2.102** 1 0.26 -6.37518 SAN STOXX 1 0.2 6.351*** 1 0.44 -0.07719 BNP STOXX 1 0.22 6.35*** 1 0.2 -6.35620 CNP STOXX 1 0.34 -0.525 1 0.2 -6.37421 COF STOXX 1 0.34 -2.344 1 0.2 -6.36322 COM STOXX 1 0.4 3.693*** 1 0.2 -6.35423 COR STOXX 1 0.6 0.627 1 0.24 -6.3824 DEU STOXX 1 0.32 6.355*** 1 0.22 -6.36625 ERS STOXX 1 0.36 0.493 1 0.34 1.903**26 FON STOXX 1 0.3 -6.355 1 0.3 0.63327 GEC STOXX 1 0.34 4.518*** 1 0.46 -0.5328 GBL STOXX 1 0.26 -2.699 1 0.38 3.204***29 SOC STOXX 1 0.2 6.351*** 1 0.2 -6.3630 HAN STOXX 1 0.32 -6.35 1 0.2 -2.98431 ICA STOXX 1 0.32 0.502 1 0.32 4.231***32 IMM STOXX 1 0.58 0.962 1 0.32 4.231***33 ING STOXX 1 0.26 6.35*** 1 0.24 -6.36334 INT STOXX 1 0.34 -4.345 1 0.28 3.23***35 KBC STOXX 1 0.24 6.351*** 1 0.2 -6.35836 KLE STOXX 1 0.36 2.593*** 1 0.36 3.947***37 MAP STOXX 1 0.48 1.519* 1 0.2 -6.36438 MED STOXX 1 0.2 6.35*** 1 0.2 -6.36439 MUE STOXX 1 0.24 -6.35 1 0.2 -6.35440 NAT STOXX 1 0.28 1.093 1 0.3 -4.9741 POH STOXX 1 0.32 6.353*** 1 0.44 0.31442 SAM STOXX 1 0.52 -0.857 1 0.2 -6.36243 SCO STOXX 1 0.32 0.759 1 0.2 -6.36444 ROD STOXX 1 0.5 -0.401 1 0.24 -6.3845 UNI STOXX 1 0.3 6.351*** 1 0.28 5.673***46 VIE STOXX 1 0.34 0.611 1 0.3 2.273***47 WEN STOXX 1 0.26 6.368*** 1 0.2 -6.35948 WER STOXX 1 0.32 -2.424 1 0.22 -6.37

Table D.4: NCoVaR from institution i on the system risk in two scenarios in period 01/01/2000till 12/31/2012 for GARCH-filtered returns. Lags determines the number of lags used in thetest. Optimal epsilon values calculated from bootstrap. T-val represents the test statistic ofNCoVaR from Eq.(10). (*),(**), (***) denotes one-sided p-value statistical significance at 10%,5% and 1%, respectively.


1 ACK STOXX 1 0.54 0.913 1 0.44 1.329*2 AEG STOXX 1 0.28 6.359*** 1 0.4 2.855***3 AGE STOXX 1 0.25 6.362*** 1 0.23 -4.6464 ALL STOXX 1 0.38 1.45* 1 0.21 -6.3665 ASS STOXX 1 0.23 6.351*** 1 0.21 -6.3596 AXA STOXX 1 0.26 6.351*** 1 0.2 -3.0357 BIR STOXX 1 0.25 -4.733 1 0.21 5.319***8 BAN STOXX 1 0.38 -1.262 1 0.23 -3.619 PAS STOXX 1 0.32 6.351*** 1 0.2 -4.38210 MIL STOXX 1 0.24 -6.369 1 0.21 -4.47711 SON STOXX 1 0.3 -3.854 1 0.31 4.083***12 ROM STOXX 1 0.27 6.351*** 1 0.27 -5.88313 BBV STOXX 1 0.25 6.352*** 1 0.21 -6.36214 POR STOXX 1 0.26 -6.351 1 0.3 -5.37815 ESS STOXX 1 0.26 -6.357 1 0.21 -5.77116 POP STOXX 1 0.29 -4.344 1 0.27 4.958***17 ESP STOXX 1 0.27 6.351*** 1 0.21 -6.36418 SAN STOXX 1 0.21 6.351*** 1 0.43 0.41819 BNP STOXX 1 0.34 -1.947 1 0.2 -3.12120 CNP STOXX 1 0.31 0.531 1 0.21 -6.3721 COF STOXX 1 0.2 -6.351 1 0.21 -6.36522 COM STOXX 1 0.26 6.351*** 1 0.2 -6.3623 COR STOXX 1 0.25 2.84*** 1 0.21 -6.37124 DEU STOXX 1 0.23 6.352*** 1 0.23 -6.3725 ERS STOXX 1 0.38 1.576* 1 0.27 3.141***26 FON STOXX 1 0.26 -6.357 1 0.32 0.05227 GEC STOXX 1 0.24 -6.362 1 0.33 -0.4628 GBL STOXX 1 0.3 -6.359 1 0.42 0.08829 SOC STOXX 1 0.24 6.351*** 1 0.22 -6.36430 HAN STOXX 1 0.27 -3.068 1 0.2 -3.91131 ICA STOXX 1 0.25 -3.335 1 0.34 3.846***32 IMM STOXX 1 0.22 -6.361 1 0.35 4.183***33 ING STOXX 1 0.26 6.351*** 1 0.22 -3.73434 INT STOXX 1 0.26 6.352*** 1 0.2 -6.35935 KBC STOXX 1 0.24 6.351*** 1 0.52 -0.31536 KLE STOXX 1 0.26 -0.444 1 0.22 -6.37137 MAP STOXX 1 0.25 1.432* 1 0.28 4.994***38 MED STOXX 1 0.27 6.351*** 1 0.2 -6.36139 MUE STOXX 1 0.29 -6.351 1 0.2 -6.35740 NAT STOXX 1 0.27 0.34 1 0.28 -4.93741 POH STOXX 1 0.29 6.352*** 1 0.24 -3.92542 SAM STOXX 1 0.59 -0.147 1 0.2 -6.36243 SCO STOXX 1 0.32 -2.487 1 0.2 -6.36544 ROD STOXX 1 0.52 0.239 1 0.23 -6.37545 UNI STOXX 1 0.28 6.352*** 1 0.3 5.886***46 VIE STOXX 1 0.32 3.747*** 1 0.2 -6.36247 WEN STOXX 1 0.25 4.869*** 1 0.22 -6.36448 WER STOXX 1 0.25 6.352*** 1 0.45 1.411*

exploring nonlinearities in financial systemic risk · exploring nonlinearities in financial...

Documents